CrystalXRD-Bench: leaderboard
CrystalXRD-Bench (Xu et al., 2026) gives a vision-language model a rendered powder X-ray diffraction pattern with the material's CIF text and formula, and asks for every Miller index contributing to the highest-intensity peak, on 250 samples from ten crystallographic databases. The score is the Jaccard index between the predicted and ground-truth index sets, macro-averaged over samples, in percent; higher is better. Models run in their strongest available reasoning mode, and unparsable answers score zero.
Metric: Jaccard Index (%; predicted vs ground-truth Miller-index set of the highest peak, macro-averaged over 250 samples). Source: arxiv.org. Saturation forecast: Around March 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (High) | 58.88 | #121 (GPT-5.4) |
| 2 | Gemini 3.1 Pro (Preview) | 49.54 | #91 |
| 3 | Gemini 3 Pro | 45.26 | #135 |
| 4 | Qwen 3.6 Plus (Thinking) | 40.49 | #217 (Qwen 3.6 Plus) |
| 5 | Kimi K2.5 (Thinking) | 21.15 | #252 (Kimi K2.5) |
| 6 | Claude Opus 4.6 (Thinking) | 20.85 | #100 (Claude Opus 4.6) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=crystalxrd-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.